abstract.tex

\section*{Abstract}
Higher harmonic generation (HHG) microscopy allows for imaging of biological tissue at subcellular resolution.
At this scale, mechanical skin properties or brain tumor presence may be observed.
Clinical settings need artificial intelligence replace time-consuming measurements and human observations.
For example, obtaining stress-strain curves from skin tissue requires mechanical measurements, and intraoperatively diagnosing tumors requires clinicians to inspect images thoroughly.
To predict stress-strain curves from HHG skin tissue images and distinguish two pediatric brain tumors, two convolutional neural networks are developed and validated.

The skin stress-strain curve predictor achieved a mean $R^2$ of -0.36 (SE 0.60).
The brain tumor model could distinguish medulloblastoma from pilocytic astrocytoma with a mean average precision of 0.89 (SE 0.05) and 0.41 (SE 0.20) AUPRG.

Both models need further training and external validation.
After additional training and validation, updated models may ultimately be used to analyze live microendoscope images to be used by plastic surgeons, or to intraoperatively discriminate between pediatric patients with pilocytic astrocytoma or medulloblastoma, or to pre-select interesting regions for diagnosis.